Gemini CLI – the real Claude Code killer?

Key Concepts:

• Gemini CLI: Google's open-source, autonomous coding agent.
• Context Window: The amount of information the AI model can consider at once (Gemini CLI boasts a 1 million context window).
• Gemini 2.5 Pro/Flash: Different versions of the Gemini model, with Pro being more powerful.
• Cloud Code: A competing coding agent.
• API Key: A method to authenticate with Gemini CLI that allows for paid usage and prevents data training by Google.
• Agents.md: A file containing optimized prompts for the coding agent.
• Settings.json: A configuration file for Gemini CLI that allows for reusing prompts.
• Vectal.ai: An AI-powered task management tool.

1. Introduction to Gemini CLI

• Google has released Gemini CLI, an open-source autonomous coding agent.
• It features a 1 million context window and offers 1,000 free queries per day.
• It is positioned as a competitor to Codex and Cloud Code.

2. Setting Up Gemini CLI

• Installation:
  • Use the command npm install -g @google/generative-ai-cli in the terminal.
  • Prerequisite: NodeJS version 18 or higher must be installed.
  • Run gemini in the terminal to launch Gemini CLI.
  • Run gemini daily to keep the CLI updated.
• Authentication:
  • Options: Login with Google, Gemini API key, or Vertex AI.
  • Login with Google: Offers 1,000 free messages per day, but Google may train on your data.
  • Gemini API Key: Requires a paid Google AI Studio account, but Google will not train on your data, and you are less likely to be downgraded to Gemini 2.5 Flash.
• Cursor Integration:
  • Open the terminal in Cursor using Command J (Mac) or Control J (Windows).
  • Type gemini to launch Gemini CLI within Cursor.
  • Choose a theme using /theme.

3. Comparison with Cloud Code

• Gemini 2.5 Pro has a 5x larger context window than Cloud Code's models (Opus and Sage).
• The speaker notes that Cloud Code has been available for longer and has had more time to mature.
• In a specific test case (improving button design), Cloud Code performed better than Gemini 2.5 Flash.

4. Configuration and Customization

• Settings.json:
  • Create a Gemini folder and a settings.json file inside it.
  • Add the following JSON structure to settings.json:

    {
      "contextFileName": "agents.md"
    }
    

  • This configuration allows Gemini CLI to use a custom agents.md file for prompts.
• Agents.md:
  • This file contains optimized prompts for the coding agent.
  • The speaker uses a 400-line prompt file originally created for Codex.

5. Building with Gemini CLI: A Practical Example

• The speaker attempts to use Gemini CLI to improve the design of thumbs up/down buttons in their AI startup, Vectal.
• The prompt: "Your task is to improve the design of the thumbs up thumbs down buttons in [file]. Instead of using these yellow emojis which are way too visible, let's use the thumbs up icons from Lucid React."
• The attempt fails due to being downgraded to Gemini 2.5 Flash.

6. Gemini CLI Tools and Features

• Web Search: Gemini CLI can browse the web using Google Search.
• Web Fetch: Gemini CLI can crawl websites.
• Memory Management: Gemini CLI has a system for managing memory using stored nodes.
• MCPS: Gemini CLI can connect and run with MCPS servers.

7. Vectal.ai: An AI-Powered Task Management Tool

• Vectal.ai is presented as a tool to increase productivity by integrating AI models into task management.
• Key features:
  • AI agents that can create and complete tasks.
  • Web research using Perplexity Pro.
  • Custom system prompts for each project.
  • Custom system prompts for each team member (on the team plan).

8. API Key Issues

• The speaker encountered issues using the Gemini API key, possibly due to high demand.
• They had to switch back to Google login to get Gemini CLI working.

9. Conclusion

• Gemini CLI has potential due to its large context window, free usage tier, and open-source nature.
• However, it needs further polishing, as Cloud Code currently performs better in certain tasks.
• The speaker encourages viewers to test Gemini CLI and provide feedback.
• If there is enough interest (indicated by "part two" comments), the speaker will create a more in-depth video on Gemini CLI.